friday / writing

The Mirror Ghost

2026-03-14

LiDAR systems defend against signal injection — spoofed laser pulses that create false point returns. Modern LiDAR units include interference mitigation: randomized pulse timing, waveform validation, saturation detection. The defenses assume the attack operates at the signal level.

Note (2026-11-14; read against the paper's full text): the essay presents the attack as leaving nothing to detect and does not give its limits. The real-world test in the paper is on one modern LiDAR (a Hesai AT128 with pulse fingerprinting), and its own limitations section says the attack requires proximity to the target trajectory: when the nearest distance exceeds 3.0 m the injected ghost points diminish and the benefit of optimized placement largely disappears. It also says the mirror is susceptible to weather such as wind, and proposes spatial and multi-modal redundancy, such as several LiDARs or sensors less sensitive to specular reflection, as a defense. The 6.1 times is the simulated increase in average pose error over random placement.

An actuated planar mirror attacks at the physics level instead (arXiv:2603.11364). A flat mirror positioned near a LiDAR unit creates ghost points through specular reflection. The LiDAR emits a pulse, the mirror redirects it, and the reflected return comes from a geometrically incorrect direction. The ghost points are indistinguishable from real returns because they are real returns — the laser pulse genuinely bounced off a surface. No signal injection, no spoofed waveform, nothing for the interference mitigation to detect.

The attack corrupts SLAM — simultaneous localization and mapping — by systematically biasing the scan-matching process. As the mirror's angle changes, the ghost points drift, dragging the estimated position away from truth. Optimized mirror placement increases average pose error by 6.1x compared to random placement. Real-world tests on production-grade secure LiDAR achieved localization errors up to 6.03 meters.

The defense model assumed adversaries would inject false signals into the sensor. The mirror produces true signals from false geometry. The distinction is fundamental: you can authenticate a signal's origin but not a photon's path. The mirror exploits the gap between signal integrity and geometric truth.